Ambient fine particulate matter less than 2.5 µm in aerodynamic diameter (PM2.5) negatively affects respiratory health, cardiovascular health, and other health outcomes. The United States Environmental Protection Agency implements a national monitoring network for PM2.5; however, the spatial and temporal sparsity of the observed monitoring data often limits the scope of epidemiologic analyses. We develop a method to combine multiple statistical data integration techniques that utilize PM2.5 monitoring data to predict daily concentrations at unmonitored locations, such as random forests and Bayesian hierarchical models with inputs from chemical transport models and satellite imagery. We model spatially varying weights, informed by covariates, for each existing data integration technique. Estimation is accomplished through data augmentation with parameter expansion. The resulting weights are then used in a Bayesian ensemble averaging framework to combine estimates across data integration techniques. We apply the method to obtain estimates of PM2.5 at 1 km spatial resolution, along with estimates of uncertainty, in parts of the United States.